simulated method of moments estimation for copula-based...
TRANSCRIPT
Simulated Method of Moments Estimation for
Copula-Based Multivariate Models�
Dong Hwan Oh and Andrew J. Patton
Duke University
30 May 2012
.�We thank Tim Bollerslev, George Tauchen, Yanqin Fan and seminar participants at Duke University,
Erasmus University Rotterdam, HECMontreal, McGill University, Monash University, Vanderbilt University,
Humboldt-Copenhagen Financial Econometrics workshop, Econometric Society Australiasian Meetings and
Econometric Society Asian Meetings for helpful comments. Contact address: Andrew Patton, Department
of Economics, Duke University, 213 Social Sciences Building, Box 90097, Durham NC 27708-0097. Email:
Simulated Method of Moments Estimation forCopula-Based Multivariate Models
Abstract
This paper considers the estimation of the parameters of a copula via a simulated method
of moments type approach. This approach is attractive when the likelihood of the copula
model is not known in closed form, or when the researcher has a set of dependence measures
or other functionals of the copula that are of particular interest. The proposed approach
naturally also nests method of moments and generalized method of moments estimators.
Drawing on results for simulation based estimation and on recent work in empirical copula
process theory, we show the consistency and asymptotic normality of the proposed estimator,
and obtain a simple test of over-identifying restrictions as a speci�cation test. The results
apply to both iid and time series data. We analyze the �nite-sample behavior of these
estimators in an extensive simulation study. We apply the model to a group of seven �nancial
stock returns and �nd evidence of statistically signi�cant tail dependence, and mild evidence
that the dependence between these assets is stronger in crashes than booms.
Keywords: correlation, dependence, inference, method of moments, SMM
J.E.L. codes: C31, C32, C51.
1
1 Introduction
Copula-based models for multivariate distributions are widely used in a variety of applica-
tions, including actuarial science and insurance (Embrechts, McNeil and Straumann 2002;
Rosenberg and Schuermann 2006), economics (Brendstrup and Paarsch 2007; Bonhomme
and Robin 2009), epidemiology (Clayton 1978; Fine and Jiang 2000), �nance (Cherubini,
Luciano and Vecchiato 2004; Patton 2006a), geology and hydrology (Cook and Johnson
1981; Genest and Favre 2007), among many others. An important bene�t they provide is
the �exibility to specify the marginal distributions separately from the dependence structure,
without imposing that they come from the same family of joint distributions.
While copulas provide a great deal of �exibility in theory, the search for copula models
that work well in practice is an ongoing one. This search has spawned a number of new and
�exible models, see Demarta and McNeil (2005), McNeil, Frey and Embrechts (2005), Smith,
Min, Almeida and Czado (2010), Smith, Gan and Kohn (2011), and Oh and Patton (2011),
among others. Some of these models are such that the likelihood of the copula is either not
known in closed form, or is complicated to obtain and maximize, motivating the considera-
tion of estimation methods other than MLE. Moreover, in many �nancial applications, the
estimated copula model is used in pricing a derivative security, such as a collateralized debt
obligation or a credit default swap (CDO or CDS), and it may be of interest to minimize
the pricing error (the observed market price less the model-implied price of the security) in
calibrating the parameters of the model. In some cases the mapping from the parameter(s)
of the copula to dependence measures (such as Spearman�s or Kendall�s rank correlation, for
example) or to the price of the derivative contract is known in closed form, thus allowing
for method of moments or generalized method of moments (GMM) estimation. In general,
however, this mapping is unknown, and an alternative estimation method is required. We
2
consider a simple yet widely applicable simulation-based approach to address this problem.
This paper presents the asymptotic properties of a simulation-based estimator of the
parameters of a copula model. We consider both iid and time series data, and we consider the
case that the marginal distributions are estimated using the empirical distribution function
(EDF). The estimation method we consider shares features with the simulated method of
moments (SMM), see McFadden (1989) and Pakes and Pollard (1989), for example, however
the presence of the EDF in the sample �moments�means that existing results on SMM are
not directly applicable. We draw on well-known results on SMM estimators, see Newey and
McFadden (1994) for example, and recent results from empirical process theory for copulas,
see Fermanian, Radulovic and Wegkamp (2004), Chen and Fan (2006) and Rémillard (2010),
to show the consistency and asymptotic normality of simulation-based estimators of copula
models. To the best of our knowledge, simulation-based estimation of copula models has
not previously been considered in the literature. An extensive simulation study veri�es that
the asymptotic results provide a good approximation in �nite samples. We illustrate the
results with an application to a model of the dependence between the equity returns on
seven �nancial �rms during the recent crisis period.
In addition to maximum likelihood, numerous other estimation methods have been con-
sidered for copula-based multivariate models. We describe these here and contrast them with
the SMM approach proposed in this paper. Multi-stage maximum likelihood, also known as
�inference functions for margins�in this literature (see Joe and Xu (1996) and Joe (2005) for
iid data and Patton (2006b) for time series data) is one of the most widely-used estimation
methods. The �maximization by parts�algorithm of Song, Fan and Kalb�eisch (2005) is an
iterative method that improves the e¢ ciency of multi-stage MLE, and attains full e¢ ciency
under some conditions. Like MLE, both of these methods only apply when the marginal
distributions are parametric. When the marginal distribution models are correctly speci�ed
3
this improves the e¢ ciency of the estimator, relative to the proposed SMM approach using
nonparametric margins, however it introduces the possibility of mis-speci�ed marginal dis-
tributions, which can have deleterious e¤ects on the copula parameter estimates, see Kim,
Silvapulle and Silvapulle (2007).
Semi-parametric maximum likelihood (see Genest, Ghoudi and Rivest (1995) for iid data
and Chen and Fan (2006), Chan, Chen, Chen, Fan and Peng (2009) and Chen, Fan and
Tsyrennikov (2006) for time series data) is also a widely-used estimation method and has a
number of attractive features. Most importantly, with respect to SMM approach proposed
here, it yields fully e¢ cient estimates of the copula parameters, whereas SMM generally
does not. Semi-parametric MLE requires, of course, the copula likelihood and for some
more complicated models the likelihood can be cumbersome to derive or to compute, e.g.
the �stochastic copula�model of Hafner and Manner (2012) or the high dimension factor
copula model of Oh and Patton (2011). In such applications it may be desirable to avoid
the likelihood and use a simpler SMM approach.
A long-standing estimator of the copula parameter is the method of moments (MM)
estimator (see Genest (1987) and Genest and Rivest (1993) for iid data and Rémillard (2010)
for time series data). This estimator exploits the known one-to-one mapping between the
parameters of certain copulas and certain measures of dependence. For example, a Clayton
copula with parameter � implies Kendall�s tau of �= (�+ 2) ; yielding a simple MM estimator
of the parameter of this copula as � = 2� = (1� �) : MM estimators usually have the bene�t
of being very fast to compute. The SMM estimator proposed in this paper is a generalization
of MM in two directions. Firstly, it allows the consideration of over-identi�ed models: For
some copulas we have more implied dependence measures than unknown parameters (e.g.,
for the Normal copula we have both Kendall�s tau and Spearman�s rank correlation in closed
form). By treating this as a GMM estimation problem we can draw on the information in
4
all available dependence measures. Secondly, we allow for dependence measures that are not
known closed-form functions of the copula parameters, and use simulations to obtain the
mapping, making this SMM rather than GMM.
Other, less-widely used, estimation methods considered in the literature include minimum
distance estimation, see Tsukahara (2005), and �expert judgment�estimation, see Britton,
Fisher and Whitley (1998). This paper contributes to this literature by considering the
properties of a SMM-type estimator, for both iid and time series data, nesting GMM and
MM estimation of the copula parameter as special cases.
2 Simulation-based estimation of copula models
We consider the same class of data generating processes (DGPs) as Chen and Fan (2006),
Chan, Chen, Chen, Fan and Peng (2009) and Rémillard (2010). This class allows each
variable to have time-varying conditional mean and conditional variance, each governed by
parametric models, with some unknown marginal distribution. As in those papers, and also
earlier papers such as Genest and Rivest (1993) and Genest, Ghoudi and Rivest (1995), we
estimate the marginal distributions using the empirical distribution function (EDF). The
conditional copula of the data is assumed to belong to a parametric family with unknown
parameter �0: The DGP we consider is:
[Y1t; : : : ; YNt]0 � Yt = �t (�0) + �t (�0)�t (1)
where �t (�) � [�1t (�) ; : : : ; �Nt (�)]0
�t (�) � diag f�1t (�) ; : : : ; �Nt (�)g
[�1t; : : : ; �Nt]0 � �t � iid F� = C (F1; : : : ; FN ;�0)
5
where �t and �t are Ft�1-measurable and independent of �t. Ft�1 is the sigma �eld contain-
ing information generated by fYt�1;Yt�2; : : :g. The r � 1 vector of parameters governing
the dynamics of the variables, �0; is assumed to bepT -consistently estimable, which holds
under mild conditions for many commonly-used models for multivariate time series, such as
ARMA models, GARCH models, stochastic volatility models, etc. If �0 is known, or if �t
and �t are known constant, then the model becomes one for iid data. Our task is to esti-
mate the p� 1 vector of copula parameters, �0 2 �; based on the (estimated) standardized
residual f�t � ��1t (�)[Yt � �t(�)]gTt=1 and simulations from the copula model, C (�;�).
2.1 De�nition of the SMM estimator
We will consider simulation from some parametric multivariate distribution, Fx (�) ; with
marginal distributions Gi (�) ; and copula C (�) : This allows us to consider cases where it
is possible to simulate directly from the copula model C (�) (in which case the Gi are all
Unif (0; 1)) and also cases where the copula model is embedded in some joint distribution
with unknown marginal distributions, such as the factor copula models of Oh and Patton
(2011).
We use only �pure� dependence measures as moments since those are a¤ected not by
changes in the marginal distributions of simulated data (X). For example, moments like
means and variances, are functions of the marginal distributions (Gi) and contain no infor-
mation on the copula. Measures like linear correlation contain information on the copula but
are also a¤ected by the marginal distributions. Dependence measures like Spearman�s rank
correlation and quantile dependence are purely functions of the copula and are una¤ected
by the marginal distributions, see Nelsen (2006) and Joe (1997) for example. Spearman�s
6
rank correlation, quantile dependence, and Kendall�s tau for the pair��i; �j
�are de�ned as:
�ij � 12E�Fi (�i)Fj
��j��� 3 = 12
Z ZuvdCij (u; v)� 3 (2)
�ijq �
8><>: P�Fi (�i) � qjFj
��j�� q�=
Cij(q;q)
q; q 2 (0; 0:5]
P�Fi (�i) > qjFj
��j�> q�=
1�2q+Cij(q;q)1�q ; q 2 (0:5; 1)
(3)
� ij � 4E�Cij�Fi (�i) ; Fj
��j���
� 1 (4)
where Cij is the copula of��i; �j
�: The sample counterparts based on the estimated stan-
dardized residuals are de�ned as:
�ij � 12
T
TXt=1
Fi (�it) Fj��jt�� 3 (5)
�ij
q �
8><>:1Tq
PTt=1 1fFi (�it) � q; Fj
��jt�� qg; q 2 (0; 0:5]
1T (1�q)
PTt=1 1fFi (�it) > q; Fj
��jt�> qg; q 2 (0:5; 1)
(6)
� ij � 4
T
TXt=1
Cij
�Fi (�it) ; Fj
��jt��� 1 (7)
where Fi (y) � (T + 1)�1PT
t=1 1f�it � yg; and Cij (u; v) � (T + 1)�1PT
t=1 1fFi (�it) �
u; Fj��jt�� vg: Counterparts based on simulations are denoted ~�ij (�), ~�ijq (�) and ~� ij (�) :
Let ~mS (�) be a (m� 1) vector of dependence measures computed using S simulations
from Fx (�), fXsgSs=1 ; and let mT be the corresponding vector of dependence measures
computed using the the standardized residuals f�tgTt=1. These vectors can also contain linear
combinations of dependence measures, a feature that is useful when considering estimation
of high-dimension models. De�ne the di¤erence between these as
gT;S (�) � mT � ~mS (�) (8)
Our SMM estimator is based on searching across � 2 � to make this di¤erence as small as
possible. The estimator is de�ned as:
7
�T;S � argmin�2�
QT;S (�) (9)
where QT;S (�) � g0T;S (�)WTgT;S (�)
and WT is some positive de�nite weight matrix, which may depend on the data. As usual,
for identi�cation we require at least as many moment conditions as there are free parameters
(i.e.,m � p). In the subsections below we establish the consistency and asymptotic normality
of this estimator, provide a consistent estimator of its asymptotic covariance matrix, and
obtain a test based on over-identifying restrictions. The supplemental appendix presents
details on the computation of the objective function.
2.2 Consistency of the SMM estimator
The estimation problem here di¤ers in two important ways from standard GMM or M-
estimation: Firstly, the objective function, QT;S (�) is not continuous in � since ~mS (�) will
be a number in a set of discrete values as � varies on �, for example,n0; 1
Sq; 2Sq; : : : ; S
Sq
ofor a lower quantile dependence. This problem would vanish if, for the copula model being
considered, we knew the mapping � 7�!m0 (�) � limS!1 ~mS (�) in closed form. The second
di¤erence is that a law of large numbers is not available to show the pointwise convergence
of gT;S (�) ; as the functions mT and ~mS (�) both involve empirical distribution functions.
We use recent developments in empirical process theory to overcome this di¢ culty.
We now list some assumptions that are required for our results to hold.
Assumption 1
(i) The distributions F� and Fx are continuous.
(ii) Every bivariate marginal copula Cij of C has continuous partial derivatives with respect
to ui and uj.
8
If the data Yt are iid; e.g. if �t and �t are known constant in equation (1), or if
�0 is known, then Assumption 1 is su¢ cient to prove Proposition 1 below, using the re-
sults of Fermanian, Radulovic and Wegkamp (2004). If, however, estimated standardized
residuals are used in the estimation of the copula then more assumptions are necessary in
order to control the estimation error coming from the models for the conditional means and
conditional variances. We combine assumptions A1�A6 in Rémillard (2010) in the follow-
ing assumption. First, de�ne 0t = ��1t
���_�t
���and 1kt = ��1t
���_�kt
���where
_�t (�) =@�t(�)@�0 ; _�kt (�) =
@[�t(�)]k-th column@�0 ; k = 1; : : : ; N: De�ne dt as
dt = �t � �t � 0t +
NXk=1
�kt 1kt
!��� �0
�where �kt is k-th row of �t and both 0t and 1kt are Ft�1-measurable.
Assumption 2
(i) 1T
TPt=1
0tp! �0 and 1
T
TPt=1
1ktp! �1k where �0 and �1k are deterministic for k =
1; : : : ; N:
(ii) 1T
TPt=1
E (k 0tk) ; 1TTPt=1
E�k 0tk
2� ; 1T
TPt=1
E (k 1ktk) ; and 1T
TPt=1
E�k 1ktk
2� are boundedfor k = 1; : : : ; N:
(iii) There exists a sequence of positive terms rt > 0 so thatP
t�1 rt <1 and such that the
sequence max1�t�T kdtk =rt is tight.
(iv) max1�t�T k 0tk =pT = op (1) and max1�t�T �kt k 1ktk =
pT = op (1) for k = 1; : : : ; N:
(v)��T ;
pT��� �0
��weakly converges to a continuous Gaussian process in [0; 1]N�Rr,
where �T is the empirical copula process of uniform random variables:
�T =1pT
TPt=1
�NQk=1
1 (Ukt � uk)� C (u)�
9
(vi) @F�@�k
and �k@F�@�k
are bounded and continuous on �RN = [�1;+1]N for k = 1; : : : ; N:
With these two assumptions, sample rank correlation and quantile dependence con-
verge in probability to their population counterparts, see Theorems 3 and 6 of Fermanian,
Radulovic and Wegkamp (2004) for the iid case, and combine with Corollary 1 of Rémillard
(2010) for the time series case. (See Lemma 1 of the supplemental appendix for details.)
When applied to simulated data this convergence holds pointwise for any �: Thus gT;S (�)
converges in probability to the population moment functions de�ned as follows:
gT;S (�) � mT � ~mS (�)p�! g0 (�) �m0 (�0)�m0 (�) ; for 8� 2 � as T; S !1 (10)
We de�ne the population objective function as
Q0 (�) = g0 (�)0W0g0 (�) (11)
whereW0 is the probability limit of WT . The convergence of gT;S (�) and WT implies that
QT;S (�)p�! Q0 (�) for 8� 2 � as T; S !1
For consistency of our estimator we need, as usual, uniform convergence of QT;S (�) ; but
as this function is not continuous in � and a law of large numbers is not available, the
standard approach based on a uniform law of large numbers is not available. We instead use
results on the stochastic equicontinuity of gT;S (�) ; based on Andrews (1994) and Newey
and McFadden (1994).
Assumption 3
(i) g0 (�) 6= 0 for � 6= �0
(ii) � is compact.
10
(iii) Every bivariate marginal copula Cij (ui;uj;�) of C (�) on (ui; uj) 2 (0; 1) � (0; 1) is
Lipschitz continuous on �:
(iv) WT is Op (1) and converges in probability toW0; a positive de�nite matrix.
Proposition 1 Suppose that Assumptions 1, 2 and 3 hold. Then �T;Sp�! �0 as T; S !1
A sketch of all proofs is presented in the Appendix, and detailed proofs are in the supple-
mental appendix. Assumption 3(iii) is needed to prove the stochastic Lipschitz continuity
of gT;S (�) ; which is a su¢ cient condition for the stochastic equicontinuity of gT;S (�), and
can easily be shown to be satis�ed for many bivariate parametric copulas. Assumption 3(ii)
requires compactness of the parameter space, a common assumption, and is aided by having
outside information (such as constraints from economic arguments) that allow the researcher
to bound the set of plausible parameters. While Pakes and Pollard (1989) and McFadden
(1989) show the consistency of SMM estimator for T; S diverging at the same rate, Proposi-
tion 1 shows that the copula parameter is consistent at any relative rate of T and S as long
as both diverge. If we know the function m (�) in closed form, then GMM is feasible and is
equivalent to our estimator with S=T !1 as T; S !1:
We focus on weak consistency of our estimator because it su¢ ces for our asymptotic distri-
bution theory, presented below. A corresponding strong consistency result, i.e., �T;Sa:s:�! �0;
may be obtained by drawing on recent work by Bouzebda and Zari (2011). The key is to
show uniform strong convergence of the sample objective function, from which strong consis-
tency of the estimator easily follows, see Newey and McFadden (1994) for example. Uniform
strong consistency of the objective function can be shown by combining minor changes in the
above assumptions (e.g., WT must converge a.s. toW0) with pointwise strong convergence
of the objective function, which can be obtained using results of Bouzebda and Zari (2011).
11
2.3 Asymptotic normality of the SMM estimator
As QT;S (�) is non-di¤erentiable the standard approach based on a Taylor expansion is not
available, however the asymptotic normality of our estimator can still be established with
some further assumptions:
Assumption 4
(i) �0 is an interior point of �
(ii) g0 (�) is di¤erentiable at �0 with derivative G0 such that G00W0G0 is nonsingular.
(iii) gT;S��T;S
�0WTgT;S
��T;S
�� inf�2� gT;S (�)0 WTgT;S (�) + op (1=T + 1=S)
The �rst assumption above is standard, and the third assumption is standard in simulation-
based estimation problems, see Newey and McFadden (1994) for example. The rate at
which the op term vanishes in part (iii) turns out to depend on the smaller of T or S; as
op (1=T + 1=S) = op�min (T; S)�1
�. The second assumption requires the population objec-
tive function, g0; to be di¤erentiable even though its �nite-sample counterpart is not, which
is common in simulation-based estimation. The nonsingularity of G00W0G0 is su¢ cient for
local identi�cation of the parameters of this model at �0; see Hall (2005) and Rothenberg
(1971). With these assumptions in hand we obtain the following result.
Proposition 2 Suppose that Assumptions 1, 2, 3 and 4 hold. Then
1p1=T + 1=S
��T;S � �0
�d�! N (0;0) as T; S !1 (12)
where 0 = (G00W0G0)
�1G00W0�0W0G0 (G
00W0G0)
�1 ; and �0 � avar [mT ] :
The rate of convergence is thus shown to equal min (T; S)1=2 : In general, one would like
to set S very large to minimize the impact of simulation error and obtain apT convergence
12
rate, however if the model is computationally costly to simulate, then the result for S �
T may be useful. When S and T diverge at di¤erent rates the asymptotic variance of
min (T; S)1=2��T;S � �0
�is simply 0:When S and T diverge at the same rate, say S=T !
k 2 (0;1) ; the asymptotic variance ofpT��T;S � �0
�is (1 + 1=k)0; which incorporates
e¢ ciency loss from simulation error. As usual we �nd that 0 =�G00�
�10 G0
��1ifW0 is the
e¢ cient weight matrix, ��10 :
The proof of the above proposition uses recent results for empirical copula processes pre-
sented in Fermanian, Radulovic and Wegkamp (2004) and Rémillard (2010) to establish the
asymptotic normality of the sample dependence measures, mT ; and requires us to establish
the stochastic equicontinuity of the moment functions, vT;S (�) =pT [gT;S (�)� g0 (�)] :
These are shown in Lemmas 6 and 7 in the supplemental appendix.
Chen and Fan (2006), Chan, Chen, Chen, Fan and Peng (2009) and Rémillard (2010)
show that estimation error from � does not enter the asymptotic distribution of the cop-
ula parameter estimator for maximum likelihood or (analytical) moment-based estimators,
and the above proposition shows that this surprising result also holds for the SMM-type
estimators proposed here. In applications based on parametric models for the marginal dis-
tributions, the asymptotic covariance matrix of the copula parameter is more complicated.
In such cases, the model is fully parametric and the estimation approach here is a form of
two-stage GMM (or SMM). In the absence of simulations, this can be treated using exist-
ing methods, see White (1994) and Gouriéroux, Monfort and Renault (1996) for example.
If simulations are used in the copula estimation step, then the lemmas presented in the
appendix can be combined with existing results on two-stage GMM to obtain the limiting
distribution. This is straightforward but requires some additional detailed notation, and so
is not presented here.
13
2.4 Consistent estimation of the asymptotic variance
The asymptotic variance of our estimator has the familiar form of standard GMM appli-
cations, however the components �0 and G0 require more care in their estimation than in
standard applications. We suggest using an iid bootstrap to estimate �0 :
1. Sample with replacement from the standardized residuals f�tgTt=1 to obtain a bootstrap
sample,n�(b)t
oTt=1. Repeat this step B times.
2. Usingn�(b)t
oTt=1; b = 1; :::; B; compute the sample moments and denote as m(b)
T ; b =
1; :::; B.
3. Calculate the sample covariance matrix of m(b)T across the bootstrap replications, and
scale it by the sample size:
�T;B =T
B
BXb=1
�m(b)T � mT
��m(b)T � mT
�0(13)
For the estimation of G0; we suggest a numerical derivative of gT;S (�) at �T;S, however
the fact that gT;S is non-di¤erentiable means that care is needed in choosing the step size
for the numerical derivative. In particular, Proposition 3 below shows that we need to let
the step size go to zero, as usual, but slower than the inverse of the rate of convergence of
the estimator (i.e., 1=min�pT ;pS�). Let ek denote the k-th unit vector whose dimension
is the same as that of �, and let "T;S denote the step size. A two-sided numerical derivative
estimator GT;S of G has k-th column
GT;S;k =gT;S
��T;S+ek"T;S
�� gT;S
��T;S�ek"T;S
�2"T;S
; k = 1; 2; :::; p (14)
Combine this estimator with WT to form:
T;S;B =�G0T;SWT GT;S
��1G0T;SWT �T;BWT GT;S
�G0T;SWT GT;S
��1(15)
14
Proposition 3 Suppose that all assumptions of Proposition 2 are satis�ed, and that
"T;S ! 0, "T;S � min�pT ;pS�! 1; B ! 1 as T; S ! 1. Then �T;B
p�! �0;
GT;Sp�! G0 and T;S;B
p�! 0 as T; S !1:
2.5 A test of overidentifying restrictions
If the number of moments used in estimation is greater than the number of copula parameters,
then it is possible to conduct a simple test of the over-identifying restrictions, which can
be used as a speci�cation test of the model. When the e¢ cient weight matrix is used in
estimation, the asymptotic distribution of this test statistic is the usual chi-squared, however
the method of proof is di¤erent as we again need to deal with the lack of di¤erentiability
of the objective function. We also consider the distribution of this test statistic for general
weight matrices, leading to a non-standard limiting distribution.
Proposition 4 Suppose that all assumptions of Proposition 3 are satis�ed and that the
number of moments (m) is greater than the number of copula parameters (p) : Then
JT;S � min (T; S)gT;S
��T;S
�0WTgT;S
��T;S
�d�! u0A0
0A0u as T; S !1
where u � N (0; I)
and A0 � W1=20 �
1=20 R0, R0� I���1=20 G0 (G
00W0G0)
�1G00W0�
1=20 : If WT = ��1T;B; then
JT;Sd�! �2m�p as usual.
As in standard applications, the above test statistic has a chi-squared limiting distribution
if the e¢ cient weight matrix (��1T;B) is used. When any other weight matrix is used, the test
statistic has a sample-speci�c limiting distribution, and critical values in such cases can be
obtained via a simple simulation:
1. Compute R using GT;S, WT , and �T;B:
15
2. Simulate u(k) s iid N (0; I), for k = 1; 2; :::; K, where K is large.
3. For each simulation, compute J (k)T;S = u(k)0R0�
1=20T;BWT �
1=2T;BRu
(k)
4. The sample (1� �) quantile ofnJ(k)T;S
oKk=1
is the critical value for this test statistic.
The need for simulations to obtain critical values from the limiting distribution is non-
standard but is not uncommon; this arises in many other testing problems, see Wolak (1989),
White (2000) and Andrews (2001) for examples. Given that u(k) is a simple standard Normal,
and that no optimization is required in this simulation, and that the matrix R need only be
computed once, obtaining critical values for this test is simple and fast.
2.6 SMM under model mis-speci�cation
All of the above results hold under the assumption that the copula model is correctly speci-
�ed. In the event that the speci�cation test proposed in the previous section rejects a model
as mis-speci�ed, one is led directly to the question of whether these results, or extensions of
them, hold for mis-speci�ed models.
In the literature on GMM, there are two common ways to de�ne mis-speci�cation. Newey
(1985) de�nes a form of �local�mis-speci�cation (where the degree of mis-speci�cation van-
ishes in the limit), and in that case it is simple to show that the asymptotic behavior of the
SMM estimator does not change at all except the mean of limit distribution. Hall and Inoue
(2003) consider �non-local�mis-speci�cation. Formally, a model is said to be mis-speci�ed
if there is no value of � 2 � which satis�es g0 (�) = 0: As Hall and Inoue (2003) note, mis-
speci�cation is only a concern when the model is over-identi�ed, and so in this section we
assume m > p: The absence of a parameter that satis�es the population moment conditions
means that we must instead consider a �pseudo-true�parameter:
16
De�nition 1 The pseudo-true parameter is �� (W0) � argmin�2� g00 (�)W0g0 (�) :
While the true parameter, �0; when it exists, is determined only by the population
moment condition g0 (�0) = 0; the pseudo-true parameter depends on the moment condition
and also on the weight matrixW0; and thus we denote it as �� (W0) : With the additional
assumptions below, the consistency of the SMM estimator under mis-speci�cation can be
proven.
Assumption 5 (i) (Non-local mis-speci�cation) kg0 (�)k > 0 for all � 2 �
(ii) (Identi�cation) There exists �� (W0) 2 � such that g0 (�� (W0))0W0g0 (�� (W0)) <
g0 (�)0W0g0 (�) for all � 2 �n f�� (W0)g
Proposition 5 Suppose Assumption 1, 2, 3(ii)-3(iv), and 5 holds. Then �T;Sp�! �� (W0)
as T; S !1
The above proposition shows that, under mis-speci�cation, the SMM estimator �T;S con-
verges in probability to the pseudo true parameter �� (W0) rather than the true parameter
�0: This extends existing results for GMM under mis-speci�cation in Hall (2000) and Hall
and Inoue (2003), as it is established under the discontinuity of the moment functions.
While consistency of �T;S under mis-speci�cation is easily obtained, establishing the limit
distribution of �T;S is not straightforward. A key contribution of Hall and Inoue (2003) is
to show that the limit distribution of GMM (with smooth, di¤erentiable moment functions)
depends on the limit distribution of the weight matrix, not merely the probability limit of
the weight matrix. In SMM applications, it is possible to show that the limit distribution
will additionally depend on the limit distribution of the numerical derivative matrix, denoted
GT;S above. Some results on the statistical properties of numerical derivatives are presented
in Hong, Mahajan and Nekipelov (2010), but this remains a relatively unexplored topic. In
17
addition to incorporating the dependence on the distribution of GT;S, under mis-speci�cation
one needs an alternative approach to establish the stochastic equicontinuity of the objective
function, which is required for a Taylor series expansion of the population objective function
to be used to obtain the limit distribution of the estimator. We leave the interesting problem
of the limit distribution of �T;S under mis-speci�cation for future research.
3 Simulation study
In this section we present a study of the �nite sample properties of the simulation-based
(SMM) estimator studied in the previous section. We consider two widely-known copula
models, the Clayton and the Gaussian (or Normal) copulas, see Nelson (2006) for discussion,
and the �factor copula�proposed in Oh and Patton (2011), outlined below. A closed-form
likelihood is available for the �rst two copulas, while the third copula requires a numerical
integration step to obtain the likelihood (details on this are presented in the supplemental
appendix). In all cases we contrast the �nite-sample properties of the MLE with the SMM
estimator. The �rst two copulas also have closed-form cumulative distribution functions,
and so quantile dependence (de�ned in equation 3) is also known in closed form. For the
Clayton copula we have Kendall�s tau in closed form (� = �= (2 + �)) but not Spearman�s
rank correlation, see Nelsen (2006). For the Normal copula we have both Spearman�s rank
correlation in closed form (�S = 6=� arcsin (�=2)) and Kendall�s tau (� = 2=� arcsin (�)) ; see
Nelsen (2006) and Demarta and McNeil (2005). This allows us to also compare GMM with
SMM for these copulas, to quantify the loss in accuracy from having to resort to simulations.
The factor copula we consider is based on the following structure:
18
Let Xi = Z + "i; i = 1; 2; :::; N
where Z s Skew t�0; �2; ��1; �
�; "i s iid t
���1�, and "i??Z 8 i (16)
[X1; :::; XN ]0 � X s Fx= C (Gx; :::; Gx)
where we use the skewed t distribution of Hansen (1994). We use the copula of X implied
by the above structure as our �factor copula�model, and it is parameterized by (�2; ��1; �) :
For the factor copula we have none of the above dependence measures in closed form, and
so simulation-based methods are required. For the simulation we set the parameters to
generate rank correlation of around 1/2, and so set the Clayton copula parameter to 1, the
Gaussian copula parameter to 1/2, and the factor copula parameters to �2 = 1; ��1 = 1=4
and � = �1=2.
We consider two di¤erent scenarios for the marginal distributions of the variables of
interest. In the �rst case we assume that the data are iid with standard Normal marginal
distributions, meaning that the only parameters that need to be estimated are those of
the copula. This simpli�ed case is contrasted with a second scenario where the marginal
distributions of the variables are assumed to follow an AR(1)-GARCH(1,1) process, which
is widely-used in time series applications:
Yit = �0 + �1Yi;t�1 + �it�it, t = 1; 2; :::; T
�2it = ! + ��2i;t�1 + ��2i;t�1�
2i;t�1 (17)
�t � [�1t; :::; �Nt]0 s iid F� = C (�;�; :::;�)
where � is the standard Normal distribution function andC can be Clayton, Gaussian, or the
factor copula implied by equation (16). We set the parameters of the marginal distributions
as [�0; �1; !; �; �] = [0:01; 0:05; 0:05; 0:85; 0:10] ; which broadly matches the values of these
parameters when estimated using daily equity return data. In this scenario the parameters
19
of the models for the conditional mean and variance are estimated, and then the estimated
standardized residuals are obtained:
�it =Yit � �0 � �1Yi;t�1
�it: (18)
These residuals are used in a second stage to estimate the copula parameters. In all cases
we consider a time series of length T = 1; 000, corresponding to approximately 4 years of
daily return data, and we use S = 25�T simulations in the computation of the dependence
measures to be matched in the SMM optimization. We use �ve dependence measures in
estimation: Spearman�s rank correlation, and the 0:05; 0:10; 0:90; 0:95 quantile dependence
measures, averaged across pairs of assets. We repeat each scenario 100 times, and in the
results below we use the identity weight matrix for estimation. (Corresponding results based
on the e¢ cient weight matrix, WT = ��1T;B; are comparable, and available in the supplemen-
tal appendix to this paper.) We also report the computation times (per simulation) for each
estimation.
Table 1 reveals that for all three dimensions (N = 2; 3 and 10) and for all three copula
models the estimated parameters are centered on the true values, with the average estimated
bias being small relative to the standard deviation, and with the median of the simulated
distribution centered on the true values. Looking across the dimension size, we see that
the copula model parameters are almost always more precisely estimated as the dimension
grows. This is intuitive, given the exchangeable nature of all three models.
Comparing the SMM estimator with the ML estimator, we see that the SMM estimators
su¤er a loss in e¢ ciency of around 50% for N = 2 and around 20% for N = 10: The
loss is greatest for the ��1 parameter of the factor copula, and is moderate and similar for
the remaining parameters. Some loss is of course expected, and this simulation indicates
that the loss is moderate overall. Comparing the SMM estimator to the GMM estimator
20
provides us with a measure of the loss in accuracy from having to estimate the population
moment function via simulation. We �nd that this loss ranges from zero to 3%, and thus
little is lost from using SMM rather than GMM. The simulation results in Table 2, where
the copula parameters are estimated after the estimation of AR(1)-GARCH(1,1) models for
the marginal distributions in a separate �rst stage, are very similar to the case when no
marginal distribution parameters are required to be estimated, consistent with Proposition
2. Thus that somewhat surprising asymptotic result is also relevant in �nite samples.
[ INSERT TABLES 1 AND 2 ABOUT HERE ]
In Table 3 we present the �nite-sample coverage probabilities of 95% con�dence intervals
based on the asymptotic normality result from Proposition 2 and the asymptotic covariance
matrix estimator presented in Proposition 3. As shown in that proposition, a critical input to
the asymptotic covariance matrix estimator is the step size used in computing the numerical
derivative matrix GT;S: This step size, "T;S; must go to zero, but at a slower rate than 1=pT :
Ignoring constants, our simulation sample size of T = 1; 000 suggests setting "T;S > 0:03;
which is much larger than standard step sizes used in computing numerical derivatives. (For
example, the default in many functions in MATLAB is a step size of around 6� 10�6, which
is an optimal choice in certain applications, see Judd (1998) for example.) We study the
impact of the choice of step size by considering a range of values from 0.0001 to 0.1. Table 3
shows that when the step size is set to 0.01 or 0.1 the �nite-sample coverage rates are close
to their nominal levels. However if the step size is chosen too small (0.001 or smaller) then
the coverage rates are much lower than nominal levels. For example, setting "T;S = 0:0001
(which is still 16 times larger than the default setting in MATLAB) we �nd coverage rates as
low as 2% for a nominal 95% con�dence interval. Thus this table shows that the asymptotic
theory provides a reliable means for obtaining con�dence intervals, so long as care is taken
21
not to set the step size too small.
Table 3 also presents the results of a study of the rejection rates for the test of over-
identifying restrictions presented in Proposition 4. Given that we consider W = I in this
table, the test statistic has a non-standard distribution, and we use K = 10; 000 simulations
to obtain critical values. In this case, the limiting distribution also depends on GT;S; and
we again compute GT;S using a step size of "T;S = 0:1; 0:01; 0:001 and 0:0001: The rejection
rates are close to their nominal levels 95% for the all three copula models.
[ INSERT TABLE 3 ABOUT HERE ]
We �nally consider the properties of the estimator under model mis-speci�cation. In
Table 4 we consider two scenarios: one where the true copula is Clayton but the model is
Normal, and one where the true copula is Normal but the model is Clayton. The pseudo-true
parameters for these two scenarios are not known in closed form, and we use a simulation of
10 million observations to estimate it. The pseudo-true parameters vary across the dimension
of the model, and we report them in the top row of each panel of Table 4. The remainder
of Table 4 has the same structure as Tables 1 and 2. Similar to those tables, in this mis-
speci�ed case we see that the estimated parameters are centered on the pseudo-true values,
with the average estimated bias being small relative to the standard deviation. These mis-
speci�ed scenarios also provide some insight into the power of the speci�cation test based on
over-identifying restrictions. We �nd that for all three dimensions and for both iid and AR-
GARCH data, the J-test rejected the null of correct speci�cation across all 100 simulations,
indicating this test has power to detect model mis-speci�cation.
These simulation results provide support for the proposed estimation method: for empir-
ically realistic parameter values and sample size, the estimator is approximately unbiased,
with estimated con�dence intervals that have coverage close to their nominal level when
22
the step size for the numerical derivative is chosen in line with our theoretical results, and
the test for model mis-speci�cation has �nite-sample rejection frequencies that are close to
their nominal levels when the model is correctly speci�ed, and has good power to reject
mis-speci�ed models.
[ INSERT TABLE 4 ABOUT HERE ]
4 Application to the dependence between �nancial �rms
This section considers models for the dependence between seven large �nancial �rms. We use
daily stock return data over the period January 2001 to December 2010, a total of T = 2515
trade days, on Bank of America, Bank of New York, Citigroup, Goldman Sachs, J.P. Morgan,
Morgan Stanley and Wells Fargo. Summary statistics for these returns are presented in
Table S4 of the supplemental appendix, and indicate that all series are positively skewed
and leptokurtotic, with kurtosis ranging from 16.0 (J.P. Morgan) to 119.8 (Morgan Stanley).
To capture the impact of time-varying conditional means and variances in each of these
series, we estimate the following autoregressive, conditionally heteroskedastic models:
rit = �0i + �1iri;t�1 + �2irm;t�1 + "it, "it = �it�it
where �2it = !i + �i�2i;t�1 + �1i"
2i;t�1 + 1i"
2i;t�1 � 1["i;t�1�0] (19)
+�2i"2m;t�1 + 2i"
2m;t�1 � 1["m;t�1�0]
where rit is the return on one of these seven �rms and rmt is the return on the S&P 500 index.
We include the lagged market index return in both the mean and variance speci�cations to
capture any in�uence of lagged information in the model for a given stock, and in the model
for the market index itself we set �1 = �1 = 1 = 0: Estimated parameters from these
models are presented in Table S5 of the supplemental appendix, and are consistent with the
23
values found in the empirical �nance literature, see Bollerslev, Engle and Nelson (1994) for
example. From these models we obtain the estimated standardized residuals, �it; which are
used in the estimation of the dependence structure.
In Table 5 we present measures of dependence between these seven �rms. The upper
panel reveals that rank correlation between their standardized residuals is 0.63 on average,
and ranges from 0.55 to 0.76. The lower panel of Table 5 presents measures of dependence
in the tails between these series. The upper triangle presents the average of the 1% and 99%
quantile dependence measures presented in equation (6), and we see substantial dependence
here, with values ranging between 0.16 and 0.40. The lower triangle presents the di¤erence
between the 90% and 10% quantile dependence measures, as a gauge of the degree of asym-
metry in the dependence structure. These di¤erences are mostly negative (14 out of 21),
indicating greater dependence during crashes than during booms.
Table 6 presents the estimation results for three di¤erent copula models of these series.
The �rst model is the well-known Clayton copula, the second is the Normal copula and the
third is a �factor copula�as proposed by Oh and Patton (2011). The �rst copula allows for
lower tail dependence, but imposes that upper tail dependence is zero. The second copula
implies zero tail dependence in both directions. The third copula allows for tail dependence
in both tails, and allows the degree of dependence to di¤er across positive and negative
realizations.
[ INSERT TABLES 5 AND 6 ABOUT HERE ]
For all three copulas we implent the SMM estimator proposed in Section 2, with the
identity weight matrix and the e¢ cient weight matrix, using �ve dependence measures:
Spearman�s rank correlation, and the 0:05; 0:10; 0:90; 0:95 quantile dependence measures,
averaged across pairs of assets. We also implement the MLE for comparison. The value
24
of the SMM objective function at the estimated parameters is presented for each model,
along with the p-value from a test of the over-identifying restrictions based on Proposition
4. We use Proposition 3 to compute the standard errors, with B = 1; 000 bootstraps used
to estimate �T;S; and "T;S = 0:1 used as the step size to compute GT;S:
The parameter estimates for the Normal and factor copula models are similar for ML
and SMM, while they are quite di¤erent for the Clayton copula. This may be explained by
the results of the test of over-identifying restrictions: the Clayton copula is strongly rejected
(with a p-value of less than 0.001 for both choices of weight matrix), while the Normal is
less strongly rejected (p-values of 0.043 and 0.001). The factor copula is not rejected using
this test for either choice of weight matrix. The improvement in �t from the factor copula
appears to come from its ability to capture tail dependence: the parameter that governs
tail dependence (��1) is signi�cantly greater than zero, while the parameter that governs
asymmetric dependence (�) is not signi�cantly di¤erent from zero.
Given that our sample period spands the �nancial crisis, one may wonder whether the
copula is constant throughout the period. To investigate this, we implement the copula
structrual break test proposed by Rémillard (2010). This test uses a Kolmogorov-Smirnov
type test statistic to compare the empirical copula before and after a given point in the
sample, and then searches across all dates in the sample. We implement this test using
1000 simulations for the �multiplier�method, and �nd a p-value of 0.001, indicating strong
evidence of a change in the copula over this period. Running this test on just the last two
years of our sample period (January 2009 to December 2010) results in a p-value of 0.191,
indicating no evidence of a change in the copula over this sub-period. We re-estimate our
three copula models using data from this sub-period, and present the results in the lower
panel of Table 6. The estimated parameters all indicate a (slight) increase in dependence in
this sub-sample relative to the full sample. Perhaps the largest change is in the � parameter
25
of the factor copula, which goes from around 8.8 (across the three estimation methods) to
around 4.4, indicating a substantial increase in the degree of tail dependence between these
�rms. The results of the speci�cation tests over this sub-sample are comparable to the full
sample results: the Clayton copula is strongly rejected, the Normal copula is rejected but
less strongly, and the factor copula is not rejected, using either weight matrix.
Figure 1 sheds some further light on the relative performance of these copula models, over
the full sample. This �gure compares the empirical quantile dependence function with those
implied by the three copula models. An iid bootstrap with B = 1; 000 replications is used to
construct pointwise con�dence intervals for the sample quantile dependence estimates. We
see here that the Clayton copula is �too asymmetric� relative to the data, while both the
Normal and the factor copula models appear to provide a reasonable �t.
[ INSERT FIGURE 1 ABOUT HERE ]
5 Conclusion
This paper presents the asymptotic properties of a new simulation-based estimator of the
parameters of a copula model, which matches measures of rank dependence implied by
the model to those observed in the data. The estimation method shares features with
the simulated method of moments (SMM), see McFadden (1989) and Newey and McFadden
(1994), for example, however the use of rank dependence measures as �moments�means that
existing results on SMM cannot be used. We extend well-known results on SMM estimators
using recent work in empirical process theory for copula estimation, see Fermanian, Radulovic
and Wegkamp (2004), Chen and Fan (2006) and Rémillard (2010), to show the consistency
and asymptotic normality of SMM-type estimators of copula models. To the best of our
knowledge, simulation-based estimation of copula models has not previously been considered
26
in the literature. We also provide a method for obtaining a consistent estimate of the
asymptotic covariance matrix, and a test of the over-identifying restrictions. Our results
apply to both iid and time series data, and an extensive simulation study veri�es that
the asymptotic results provide a good approximation in �nite samples. We illustrate the
results with an application to a model of the dependence between the equity returns on
seven �nancial �rms, and �nd evidence of statistically signi�cant tail dependence, and some
evidence that the dependence between these assets is stronger in crashes than booms.
Appendix: Sketch of proofsDetailed proofs are available in the supplemental appendix to this paper.
Proof of Proposition 1. First note that: (a) Q0 (�) is uniquely minimized at �0 by
Assumption 3(i) and positive de�niteW0 of Assumption 3(iv); (b)� is compact by Assump-
tion 3(ii); (c) Q0 (�) consists of linear combinations of rank correlations and quantile depen-
dence measures that are functions of pair-wise copula functions, so Q0 (�) is continuous by
Assumption 3(iii). The main part of the proof requires establishing that QT;S uniformly con-
verges in probability to Q0; which we show using �ve lemmas in the supplemental appendix:
Pointwise convergence of gT;S (�) to g0 (�) and stochastic Lipschitz continuity of gT;S (�) is
shown using results from Fermanian, Wegkamp and Radulovic (2004) and Rémillard (2010),
combined with Assumption 3(iii). This is su¢ cient for the stochastic equicontinuity of gT;S
and for the uniform convergence in probability of gT;S to g0 by Lemmas 2.8 and 2.9 of Newey
and McFadden (1994). Using the triangle and Cauchy-Schwarz inequalities this implies that
QT;S uniformly converges in probability to Q0. We have thus veri�ed that the conditions of
Theorem 2.1 of Newey and McFadden (1994) hold, and we have �p! �0 as claimed.
27
Proof of Proposition 2. We prove this proposition by verifying the �ve conditions of
Theorem 7.2 of Newey and McFadden (1994) for our problem: (i) g0 (�0) = 0 by construction
of g0 (�) = m (�0) �m (�). (ii) g0 (�) is di¤erentiable at �0 with derivative G0 such that
G00W0G0 is nonsingular by Assumption 4(ii). (iii) �0 is an interior point of� by Assumption
4(i). (iv) This part requires showing the asymptotic normality ofpTgT;S (�0) : We will
present the result only for S=T ! k 2 (0;1) : The results for the cases that S=T !
0 or S=T ! 1 are similar. In Lemma 6 of the supplemental appendix we show thatpT (mT �m0 (�0))
d! N (0;�0) as T ! 1 andpS ( ~mS (�0)�m0 (�0))
d! N (0;�0) as
S !1 using Theorem 3 and Theorem 6 of Fermanian, Radulovic and Wegkamp (2004) and
Corollary 1, Proposition 2 and Proposition 4 of Rémillard (2010). This implies that
pTgT;S (�0) =
pT (mT �m0 (�0))| {z }
d!N(0;�0)
�rT
S|{z}!1=
pk
pS ( ~mS (�0)�m0 (�0))| {z }
d!N(0;�0)
and sopTgT;S (�0)
d! N (0; (1 + 1=k)�0) as T; S !1: (v) This part requires showing that
supk���0k<�pT gT;S (�)� gT;S (�0)� g0 (�) = h1 +pT k� � �0ki p! 0: The main part of
this proof involves showing the stochastic equicontinuity of vT;S (�) =pT�gT;S (�)� g0 (�)
�:
This is shown in Lemma 7 of the supplemental appendix by showing that fg�;� (�) : � 2 �g
is a type II class of functions in Andrews (1994), and then using that paper�s Theorem 1.
Proof of Proposition 3. If �t and �t are known constant, or if �0 is known,
then the consistency of �T;B follows from Theorems 5 and 6 of Fermanian, Radulovic and
Wegkamp (2004). When �0 is estimated, the result is obtained by combining the results
in Fermanian, Radulovic and Wegkamp with those of Rémillard (2010): For simplicity,
assume that only one dependence measure is used. Let �ij and �(b)ij be the sample rank
correlations constructed from the standardized residuals��it; �
jt
Tt=1
and from the bootstrap
counterpartn�(b)it ; �
(b)jt
oTt=1: Also, de�ne the corresponding estimates, ��ij and ��
(b)ij ; using the
28
true innovations��it; �
jt
Tt=1and the bootstrapped true innovations
n�(b)it ; �
(b)jt
oTt=1(where the
same bootstrap time indices are used for bothn�(b)it ; �
(b)jt
oTt=1
andn�(b)it ; �
(b)jt
oTt=1). De�ne
true � as �0: Theorem 5 of Fermanian, Radulovic and Wegkamp (2004) shows that
pT���ij � �0
�=pT���(b)ij � ��ij
�+ op (1)
Corollary 1 and Proposition 4 of Rémillard (2010) shows, under Assumption 2, that
pT��ij � ��ij
�= op (1) and
pT��(b)ij � ��
(b)ij
�= op (1)
Combining those three equations, we obtain
pT��ij � �0
�=pT��(b)ij � �ij
�+ op (1) , as T;B !1
and so equation (13) is a consistent estimator of �0: Consistency of the numerical derivatives
GT;S can be established using a similar approach to Theorem 7.4 of Newey and McFadden
(1994), and since WTp!W0 by Assumption 3(iv), we thus have T;S;B
p! 0:
Proof of Proposition 4. We consider only the case where S=T !1 or S=T ! k > 0.
The case for k = 0 is analogous. A Taylor expansion of g0��T;S
�around �0 yields
pTg0
��T;S
�=pTg0 (�0) +G0 �
pT��T;S��0
�+ o
�pT �T;S��0 �
and since g0 (�0) = 0 andpT �T;S��0 = Op (1)
pTg0
��T;S
�= G0 �
pT��T;S��0
�+ op (1) (20)
Then consider the following expansion of gT;S��T;S
�around �0
pTgT;S
��T;S
�=pTgT;S (�0) + GT;S �
pT��T;S��0
�+RT;S
��T;S
�(21)
where the remaining term is captured by RT;S
��T;S
�: Combining equations (20) and (21)
we obtain
pThgT;S
��T;S
�� gT;S (�0)� g0
��T;S
�i=�GT;S�G0
��pT��T;S��0
�+RT;S
��T;S
�+op (1)
29
The stochastic equicontinuity of vT;S (�) =pT [gT;S (�)� g0 (�)] is established in the proof
of Proposition 2, which implies (see proof of Proposition 2) that
pThgT;S
��T;S
�� gT;S (�0)� g0
��T;S
�i= op (1)
By Proposition 3, GT;S�G0 = op (1) ; which implies RT;S
��T;S
�= op (1) : Thus, we obtain
the expansion of gT;S��T;S
�around �0 :
pTgT;S
��T;S
�=pTgT;S (�0) + GT;S �
pT��T;S��0
�+ op (1) (22)
The remainder of the proof is the same as in standard GMM applications, see Hall (2005)
for example.
Proof of Proposition 5. Lemma 1, 2, 3 and 4 are not a¤ected by mis-speci�cation.
Lemma 5 (i) is replaced by Assumption 5 (ii). Therefore, �T;Sp! �� (W0) :
References
[1] Andrews, D.W.K., 1994, Empirical Process Methods in Econometrics, in R. Engle andD.McFadden (eds.), Handbook of Econometrics, 4, 2247-2294.
[2] Andrews, D.W.K., 2001, Testing When a Parameter is on the Boundary of the Main-tained Hypothesis, Econometrica, 69, 683-734.
[3] Bollerslev, T., R.F. Engle, and D.B. Nelson, 1994, ARCH Models, in in R.F. Engle andD. McFadden (eds.), Handbook of Econometrics, 4, 2959-3038.
[4] Bonhomme, S. and J.-M. Robin, 2009, Assessing the Equalizing Force of Mobility UsingShort Panels: France, 1990-2000, Review of Economic Studies, 76, 63-92.
[5] Bouzebda, S. and T. Zari, 2011, Strong Approximation of Empirical Copula Processesby Gaussian Processes, preprint available at http://arxiv.org/pdf/0811.3330.pdf
[6] Brendstrup, B. and H.J. Paarsch, 2007, Semiparametric Identi�cation and Estimationin Multi-Object English Auctions, Journal of Econometrics, 141, 84-108.
[7] Britton, E., P. Fisher, and J. Whitley, 1998, The In�ation Report Projections: Under-standing the Fan Chart, Bank of England Quarterly Bulletin, 30-37.
30
[8] Chen, X. and Y. Fan, 2006, Estimation and Model Selection of Semiparametric Copula-based Multivariate Dynamic Models under Copula Misspeci�cation, Journal of Econo-metrics, 135, 125-154.
[9] Chan, N.-H., J. Chen, X. Chen, Y. Fan and L. Peng, 2009, Statistical Inference forMultivariate Residual Copula of GARCH Models, Statistica Sinica, 19, 53-70.
[10] Chen, X., Y. Fan, and V. Tsyrennikov, 2006, E¢ cient Estimation of SemiparametricMultivariate Copula Models, Journal of the American Statistical Association, 101, 1228-1240.
[11] Cherubini, U., E. Luciano, and W. Vecchiato, 2004, Copula Methods in Finance, JohnWiley & Sons, England.
[12] Clayton, D.G., 1978, A model for association in bivariate life tables and its applicationin epidemiological studies of familial tendency in chronic disease incidence, Biometrika,65, 141-151.
[13] Cook, R.D. and M.E Johnson, 1981, A family of multivariate distributions for modellingnon-elliptically symmetric multivariate data, Journal of the Royal Statistical Society �B, 43, 210-218.
[14] Demarta, S. and A.J. McNeil, 2005, The t Copula and Related Copulas, InternationalStatistical Review, 73, 111-129.
[15] Embrechts, P., A. McNeil, and D. Straumann, 2002, Correlation and Dependence Prop-erties in Risk Management: Properties and Pitfalls, in M. Dempster, ed., Risk Manage-ment: Value at Risk and Beyond, Cambridge University Press.
[16] Fermanian, J., D. Radulovic, and M. Wegkamp, 2004, Weak Convergence of EmpiricalCopula Process, Bernoulli , 10, 847-860.
[17] Fine, J.P. and H. Jiang, 2000, On association in a copula with time transformations,Biometrika, 87, 559-571.
[18] Genest, C., 1987, Frank�s Family of Bivariate Distributions, Biometrika, 74, 549-555.
[19] Genest, C. and L.-P. Rivest, 1993, Statistical Inference Procedures for BivariateArchimedean Copulas, Journal of the American Statistical Association, 88, 1034-1043.
[20] Genest, C. and A.-C. Favre, 2007, Everything you always wanted to know about copulamodeling but were afraid to ask, Journal of Hydrologic Engineering, 12, 347-368.
[21] Genest, C., K. Ghoudi, and L.-P. Rivest, 1995, A Semiparametric Estimation Procedureof Dependence Parameters in Multivariate Families of Distributions, Biometrika, 82,543-552.
31
[22] Gouriéroux, C., A. Monfort, and E. Renault, 1996, Two-stage generalized momentmethod with applications to regressions with heteroscedasticity of unknown form, Jour-nal of Statistical Planning and Inference, 50, 37-63.
[23] Hafner, C. M. and H. Manner, 2012, Dynamic Stochastic Copula Models: Estimation,Inference and Applications, Journal of Applied Econometrics, 27, 269-295.
[24] Hall, A.R., 2000, Covariance matrix estimation and the power of the overidentifyingrestrictions test, Econometrica, 68, 1517-1528.
[25] Hall, A.R., 2005, Generalized Method of Moments, Oxford University Press, Oxford.
[26] Hall, A.R. and A. Inoue, 2003, The large sample behaviour of the generalized methodof moments estimator in misspeci�ed models, Journal of Econometrics, 114, 361-394.
[27] Hansen, B.E., 1994, Autoregressive Conditional Density Estimation, International Eco-nomic Review, 35, 705-730.
[28] Hong, H., A. Mahajan and D. Nekipelov, 2010, Extremum Estimation and NumericalDerivatives, working paper, Stanford University.
[29] Joe, H., 1997, Multivariate Models and Dependence Concepts, Monographs in Statisticsand Probability 73, Chapman and Hall, London.
[30] Joe, H., 2005, Asymptotic e¢ ciency of the two-stage estimation method for copula-based models, Journal of Multivariate Analysis, 94, 401-419.
[31] Joe, H. and J.J Xu, 1996, The Estimation Method of Inference Functions for Marginsfor Multivariate Models. working paper, Department of Statistics, University of BritishColumbia.
[32] Judd, K.L., 1998, Numerical Methods in Economics, MIT Press, Cambridge, USA.
[33] Kim, G., M. J. Silvapulle and P. Silvapulle, 2007, Comparison of Semiparametric andParametric Methods for Estimating Copulas, Computational Statistics & Data Analysis,51, 2836-2850.
[34] McFadden, D., 1989, A Method of Simulated Moments for Estimation of Discrete Re-sponse Models without Numerical Integration, Econometrica, 47, 995-1026.
[35] McNeil, A.J., R. Frey, and P. Embrechts, 2005, Quantitative Risk Management, Prince-ton University Press, New Jersey.
[36] Nelsen, R.B., 2006, An Introduction to Copulas, Second Edition, Springer, U.S.A.
[37] Newey, W.K., 1985, Generalized Method of Moments Speci�cation Testing, Journal ofEconometrics, 29, 229-256.
32
[38] Newey, W.K. and D. McFadden, 1994, Large Sample Estimation and Hypothesis Test-ing, in R.Engle and D. McFadden (eds.), Handbook of Econometrics, 4, 2111-2245.
[39] Oh, D.-H. and A.J. Patton, 2011, Modelling Dependence in High Dimensions usingFactor Copulas, working paper, Duke University.
[40] Pakes, A. and D. Pollard, 1989, Simulation and the Asymptotics of Optimization Esti-mators, Econometrica, 47, 1027-1057.
[41] Patton, A.J., 2006a, Modelling Asymmetric Exchange Rate Dependence, InternationalEconomic Review, 47, 527-556.
[42] Patton, A.J., 2006b, Estimation of Multivariate Models for Time Series of PossiblyDi¤erent Lengths, Journal of Applied Econometrics, 21, 147-173.
[43] Rémillard, B., 2010, Goodness-of-�t Tests for Copulas of Multivariate Time Series,working paper.
[44] Rosenberg, J.V. and T. Schuermann, 2006, A general approach to integrated risk man-agement with skewed, fat-tailed risks, Journal of Financial Economics, 79, 569-614.
[45] Rothenberg, T.J., 1971, Identi�cation of Parametric Models, Econometrica, 39, 577-591.
[46] Smith, M., A. Min, C. Almeida, and C. Czado, 2010, Modeling longitudinal data usinga pair-copula decomposition of serial dependence, Journal of the American StatisticalAssociation, 105, 1467-1479.
[47] Smith, M., Q. Gan, and R. Kohn, 2011, Modeling Dependence using Skew t Copulas:Bayesian Inference and Applications, Journal of Applied Econometrics, forthcoming.
[48] Song, P., Y. Fan and J. Kalb�eisch, 2005, Maximization by Parts in Likelihood Inference,Journal of the American Statistical Association, 100, 1145-1158.
[49] Tsukahara, H., 2005, Semiparametric estimation in copula models, Canadian Journalof Statistics, 33, 357-375.
[50] White, H., 1994, Estimation, Inference and Speci�cation Analysis, Econometric SocietyMonographs No. 22, Cambridge University Press, Cambridge, U.K.
[51] White, H., 2000, A Reality Check for Data Snooping, Econometrica, 68, 1097-1126.
[52] Wolak, F.A., 1989, Testing Inequality Contraints in Linear Econometic Models, Journalof Econometrics, 41, 205-235.
33
Table1:Simulation
resultsforiiddata
Clayton
Normal
Factorcopula
MLE
GMM
SMM
SMM�
MLE
GMM
SMM
MLE
SMM
��
��
��
��2
��1
��2
��1
�
True
1.00
1.00
1.00
1.00
0.5
0.5
0.5
1.00
0.25
-0.50
1.00
0.25
-0.50
N=2
Bias
0.001-0.014
-0.006
-0.004
0.004-0.001
-0.001
0.026-0.002
-0.026
0.016-0.012
-0.089
Stdev
0.085
0.119
0.122
0.110
0.024
0.034
0.034
0.135
0.045
0.144
0.152
0.123
0.199
Median
1.011
0.982
0.991
0.998
0.503
0.497
0.498
1.027
0.251
-0.500
0.985
0.243-0.548
90-10%
0.216
0.308
0.293
0.294
0.063
0.086
0.087
0.331
0.118
0.355
0.374
0.363
0.540
Time
0.061
0.060
512
49.5
0.021
0.292
0.483
254
103
N=3
Bias
0.015
0.008
0.006
0.008
0.003-0.004
-0.005
0.014-0.001
-0.012
0.032-0.002
-0.057
Stdev
0.061
0.090
0.092
0.091
0.020
0.025
0.026
0.120
0.028
0.109
0.124
0.111
0.157
Median
1.013
1.003
0.999
0.998
0.503
0.497
0.499
1.001
0.250
-0.502
1.031
0.256-0.542
90-10%
0.155
0.226
0.219
0.216
0.049
0.064
0.068
0.297
0.073
0.222
0.297
0.293
0.395
Time
0.113
0.091
1360
56.2
0.023
0.293
0.815
263
136
N=10
Bias
0.008
0.007
0.008
0.004
0.003-0.002
-0.002
0.011
0.000
0.006
0.026
0.001-0.011
Stdev
0.050
0.068
0.066
0.059
0.014
0.017
0.017
0.092
0.016
0.063
0.093
0.070
0.082
Median
1.005
1.002
1.005
0.999
0.504
0.498
0.499
1.005
0.248
-0.494
1.013
0.255-0.508
90-10%
0.132
0.198
0.177
0.152
0.035
0.039
0.045
0.240
0.034
0.166
0.248
0.186
0.168
Time
0.409
0.998
22289
170
0.475
0.331
3.140
396
341
34
Table2:Simulation
resultsforAR-GARCHdata
Clayton
Normal
Factorcopula
MLE
GMM
SMM
SMM�
MLE
GMM
SMM
MLE
SMM
��
��
��
��2
��1
��2
��1
�
True
1.00
1.00
1.00
1.00
0.5
0.5
0.5
1.00
0.25
-0.50
1.00
0.25
-0.50
N=2
Bias
-0.005
-0.029
-0.028
-0.020
0.003-0.001
-0.001
0.021-0.009
-0.029
0.015-0.012
-0.073
Stdev
0.087
0.124
0.124
0.108
0.024
0.035
0.036
0.137
0.046
0.150
0.155
0.121
0.188
Median
0.998
0.977
0.975
0.982
0.503
0.497
0.499
1.021
0.245-0.503
0.995
0.235-0.558
90-10%
0.228
0.327
0.340
0.267
0.061
0.084
0.090
0.343
0.118
0.382
0.411
0.346
0.509
Time
0.026
0.059
525
520.030
0.299
0.505
234
95
N=3
Bias
0.006-0.007
0.002
-0.008
0.003-0.005
-0.006
0.007-0.007
-0.011
0.013-0.020
-0.052
Stdev
0.060
0.087
0.088
0.080
0.020
0.026
0.026
0.118
0.028
0.110
0.121
0.106
0.148
Median
1.005
0.991
0.994
0.981
0.502
0.497
0.499
0.997
0.243-0.502
1.005
0.238-0.521
90-10%
0.145
0.205
0.213
0.195
0.050
0.065
0.068
0.315
0.074
0.224
0.311
0.297
0.357
Time
0.127
0.108
1577
730.022
0.288
1.009
232
119
N=10
Bias
-0.004
-0.002
-0.004
-0.003
0.002-0.003
-0.004
0.005-0.006
0.008
-0.004
-0.016
-0.012
Stdev
0.049
0.067
0.064
0.059
0.014
0.016
0.017
0.091
0.015
0.063
0.085
0.079
0.071
Median
0.995
0.996
0.987
0.988
0.503
0.497
0.497
1.002
0.243-0.493
0.990
0.238-0.508
90-10%
0.134
0.179
0.170
0.152
0.034
0.041
0.045
0.240
0.037
0.169
0.210
0.209
0.165
Time
0.292
1.059
24549
171
1.099
0.392
3.437
430
309
35
NotestoTable1:Thistablepresentstheresultsfrom
100simulationsofClayton,theNormalcopula,andafac-
torcopula.IntheSMMandGMMestimationallthreecopulasuse�vedependencemeasures,includingfourquantile
dependencemeasures(q=0:05;0:10;0:90;0;95).TheNormalandfactorcopulasalsouseSpearman�srankcorrelation,
whiletheClaytoncopulauseseitherKendall�s(GMMandSMM)orSpearman�s(SMM� )rankcorrelation.Themarginal
distributionsofthedataareassumedtobeiidN(0;1).ProblemsofdimensionN=2;3and10areconsidered,the
samplesizeisT=1;000andthenumberofsimulationsusedisS=25�T:The�rstrowofeachpanelpresentsthe
averagedi¤erencebetweentheestimatedparameteranditstruevalue.Thesecondrowpresentsthestandarddeviation
oftheestimatedparameters.Thethirdandfourthrowspresentthemedianandthedi¤erencebetweenthe90thand10th
percentilesofthedistributionofestimatedparameters.Thelastrowineachpanelpresentstheaveragetimeinsecondsto
computetheestimator.
NotestoTable2:Thistablepresentstheresultsfrom
100simulationsofClayton,theNormalcopula,andafac-
torcopula.IntheSMMandGMMestimationallthreecopulasuse�vedependencemeasures,includingfourquantile
dependencemeasures(q=0:05;0:10;0:90;0;95).TheNormalandfactorcopulasalsouseSpearman�srankcorrelation,
whiletheClaytoncopulauseseitherKendall�s(GMMandSMM)orSpearman�s(SMM� )rankcorrelation.Themarginal
distributionsofthedataareassumedtofollowAR(1)-GARCH(1,1)processes,asdescribedinSection3.
Problemsof
dimensionN=2;3and10areconsidered,thesamplesizeisT=1;000andthenumberofsimulationsusedisS=25�T:
The�rstrowofeachpanelpresentstheaveragedi¤erencebetweentheestimatedparameteranditstruevalue.Thesecond
rowpresentsthestandarddeviationoftheestimatedparameters.Thethirdandfourthrowspresentthemedianandthe
di¤erencebetweenthe90thand10thpercentilesofthedistributionofestimatedparameters.Thelastrowineachpanel
presentstheaveragetimeinsecondstocomputetheestimator.
36
Table 3: Simulation results on coverage rates
Clayton Normal Factor copula
� J � J �2 ��1 � J
N = 2"T;S0.1 91 98 94 98 94 100 95 980.01 46 99 92 98 94 99 96 1000.001 2 99 76 98 76 79 74 990.0001 1 99 21 98 54 75 57 97
N = 3"T;S0.1 97 99 89 97 99 100 96 990.01 63 98 88 97 99 96 95 1000.001 11 98 83 98 92 84 93 1000.0001 2 100 38 99 57 70 61 99
N = 10"T;S0.1 96 99 87 97 97 97 95 980.01 88 99 87 96 96 97 97 970.001 18 100 87 98 97 95 88 970.0001 0 98 71 97 73 85 81 98
Notes: This table presents the results from 100 simulations of Clayton copula, the Normalcopula, and a factor copula, all estimated by SMM. The marginal distributions of the dataare assumed to follow AR(1)-GARCH(1,1) processes, as described in Section 3. Problems ofdimension N = 2; 3 and 10 are considered, the sample size is T = 1; 000 and the number ofsimulations used is S = 25 � T: The rows of each panel contain the step size, "T;S; used incomputing the matrix of numerical derivatives, GT;S: The numbers in column �; �; �2; ��1;and � present the percentage of simulations for which the 95% con�dence interval based onthe estimated covariance matrix contained the true parameter. The numbers in column Jpresent the percentage of simulations for which the test statistic of over-identifying restric-tions test described in Section 2 was smaller than its computed critical value under 95%con�dence level.
37
Table 4: Simulation results for mis-speci�ed models
iid AR-GARCH
True copula Normal Clayton Normal ClaytonModel Clayton Normal Clayton Normal
N = 2Pseudo-true 0.542 0.599 0.543 0.588Bias -0.013 0.111 -0.007 0.046St dev 0.050 0.173 0.035 0.120Median 0.526 0.659 0.539 0.61790-10% 0.130 0.433 0.091 0.265Time 4 72 1 70J test prob. 0 0 0 0
N = 3Pseudo-true 0.543 0.599 0.542 0.607Bias 0.003 0.077 -0.002 0.006St dev 0.039 0.164 0.027 0.088Median 0.544 0.629 0.540 0.60990-10% 0.107 0.432 0.072 0.198Time 5 90 1 86J test prob. 0 0 0 0
N = 10Pseudo-true 0.544 0.602 0.544 0.603Bias 0.001 0.059 -0.001 0.047St dev 0.033 0.118 0.016 0.116Median 0.546 0.622 0.540 0.61890-10% 0.086 0.307 0.043 0.314Time 20 206 4 207J test prob. 0 0 0 0
Notes: This table presents the results from 100 simulations when the true copula andthe model are di¤erent (i.e., the model is mis-speci�ed). The parameters of the copulamodels are estimated using SMM based on rank correlation and four quantile dependencemeasures (q = 0:05; 0:10; 0:90; 0; 95). The marginal distributions of the data are assumed tobe either iid N (0; 1) or AR(1)-GARCH(1,1) processes, as described in Section 3. Problemsof dimension N = 2; 3 and 10 are considered, the sample size is T = 1; 000 and the numberof simulations used is S = 25� T: The pseudo-true parameter is estimated using 10 millionobservations. The last row in each panel presents the proportion of tests of over-identifyingrestrictions that are smaller than the 95% critical value.
38
Table 5: Sample dependence statistics
Bank of Bank of Citi Goldman JP Morgan WellsAmerica N.Y. Group Sachs Morgan Stanley Fargo
BoA 0.586 0.691 0.556 0.705 0.602 0.701BoNY 0.551 0.574 0.578 0.658 0.592 0.595Citi 0.685 0.558 0.608 0.684 0.649 0.626Goldman 0.564 0.565 0.609 0.655 0.759 0.548JPM 0.713 0.633 0.694 0.666 0.667 0.683Morgan S 0.604 0.587 0.650 0.774 0.676 0.578Wells F 0.715 0.593 0.636 0.554 0.704 0.587
BoA 0.219 0.239 0.219 0.398 0.298 0.358BoNY -0.048 0.179 0.199 0.159 0.219 0.199Citi -0.045 -0.004 0.199 0.318 0.219 0.199Goldman -0.068 0.000 0.032 0.239 0.378 0.199JPM -0.024 -0.056 -0.012 0.012 0.239 0.358Morgan S -0.060 -0.020 -0.064 -0.036 -0.008 0.219Wells F 0.020 -0.052 0.044 -0.028 0.024 0.000
Notes: This table presents measures of dependence between the seven �nancial �rms underanalysis. The upper panel presents Spearman�s rank correlation (upper triangle) and linearcorrelation (lower triangle), and the lower panel presents the di¤erence between the 10%tail dependence measures (lower triangle) and average 1% upper and lower tail dependence(upper triangle). All dependence measures are computed using the standardized residualsfrom the models for the conditional mean and variance.
39
Table6:Estimationresultsfordailyreturnson
sevenstocks
PanelA:Fullsample(2001-2010)
Clayton
Normal
Factor
MLE
SMM
SMM-opt
MLE
SMM
SMM-opt
MLE
SMM
SMM-opt
��
��
��
�2;��1;�
�2;��1;�
�2;��1;�
Estimate
0.907
1.274
1.346
0.650
0.682
0.659
1.995
2.019
1.955
Stderr
0.028
0.048
0.037
0.007
0.010
0.008
0.020
0.077
0.069
Estimate
��
��
��
0.159
0.088
0.115
Stderr
��
��
��
0.010
0.034
0.033
Estimate
��
��
��
-0.021
-0.015
-0.013
Stderr
��
��
��
0.032
0.035
0.034
QSMM�100
�19.820
19.872
�0.240
0.719
�0.040
0.187
Jpval
�0.000
0.000
�0.043
0.001
�0.139
0.096
Time
0.7
344
360
0.5
66
1734
801
858
40
PanelB:Sub-sample(2009-2010)
Clayton
Normal
Factor
MLE
SMM
SMM-opt
MLE
SMM
SMM-opt
MLE
SMM
SMM-opt
��
��
��
�2;��1;�
�2;��1;�
�2;��1;�
Estimate
1.053
1.649
1.581
0.660
0.768
0.701
2.165
2.567
2.575
Stderr
0.072
0.152
0.094
0.015
0.370
0.192
0.018
0.249
0.256
Estimate
��
��
��
0.167
0.266
0.297
Stderr
��
��
��
0.062
0.081
0.078
Estimate
��
��
��
-0.095
0.033
0.050
Stderr
��
��
��
0.062
0.070
0.067
QSMM�100
�25.295
29.702
�1.198
4.092
�0.039
0.152
Jpval
�0.000
0.000
�0.055
0.000
�0.748
0.858
Time
0.5
5357
0.4
11.4
316
150
156
Notes:Thistablepresentsestimationresultsforvariouscopulamodelsappliedtosevendaily
stockreturnsinthe
�nancialsectorovertheperiodJanuary2001toDecember2010(PanelA)andJanuary2009toDecember2010(PanelB).
Estimatesandasymptoticstandarderrorsforthecopulamodelparametersarepresented,aswellasthevalueoftheSMM
objectivefunctionattheestimatedparametersandthep-valueoftheoveridentifyingrestrictiontest.Estimateslabeled
�SMM�areestimatedusingtheidentityweightmatrix;estimateslabeled�SMM-opt�areestimatedusingthee¢cient
weightmatrix.
41
Figure 1: This �gure plots the probability of both variables being less than their q quantile(q<0.5) or greater than the q quantile (q>0.5). For the data this is averaged across all pairs,and a bootstrap 90% (pointwise) con�dence interval is presented.
.
42